Pré-Publication, Document De Travail Année : 2025

SIMuLDiTex: A Single Image Multiscale and Lightweight Diffusion Model for Texture Synthesis

SIMuLDiTex: Un modèle de diffusion léger, mono-image et multi-échelle pour la synthèse de texture.

Résumé

We introduce SIMuLDiTex, a Single Image Multi-scale & Light-weight Diffusion Texture model. Recent generative models, though effective, suffer from high computational costs and memory requirements. Our approach addresses these challenges by employing a coarse-to-fine strategy that accelerates sampling and maintains high-resolution fidelity without the need for pre-trained auto-encoders. We introduce a scale-conditioning mechanism that ensures spatial consistency across scales in the diffusion process using a compact U-Net with only one million parameters. Experiments demonstrate that SIMuLDiTex outperforms existing methods in speed, quality, and scalability, offering a practical solution for interactive-time and high-resolution texture generation. Code and models are available at https: //github.com/PierrickCh/SIMuLDiTex.

Fichier principal
Vignette du fichier
ICCV2025_arxiv-3.pdf (118.63 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
licence

Dates et versions

hal-04994907 , version 1 (18-03-2025)

Licence

Identifiants

  • HAL Id : hal-04994907 , version 1

Citer

Pierrick Chatillon, Julien Rabin, David Tschumperlé. SIMuLDiTex: A Single Image Multiscale and Lightweight Diffusion Model for Texture Synthesis. 2025. ⟨hal-04994907⟩
406 Consultations
22 Téléchargements

Partager

  • More